Open Access

ARTICLE

SegTSF: Hierarchical Segment Learning For Lightweight Multivariate Time-Series ForeCasting

Hyunjun Park¹, Hee-Gook Jun², Seongyong Kim³, Dong-Hyuk Im^4,*

1 Department of Artificial Intelligence Application, Kwangwoon University, Seoul, Republic of Korea
2 Finda, Seoul, Republic of Korea
3 Division of Big Data AI, Hoseo University, Asan, Republic of Korea
4 School of Information Convergence, Kwangwoon University, Seoul, Republic of Korea

* Corresponding Author: Dong-Hyuk Im. Email:

Computer Modeling in Engineering & Sciences 2026, 147(3), 36 https://doi.org/10.32604/cmes.2026.082506

Received 17 March 2026; Accepted 09 June 2026; Issue published 30 June 2026

Abstract

Time-series forecasting can significantly aid decision-making in fields in which immediate action is required, such as power demand forecasting, financial market analysis, and traffic flow management. Transformer-based models achieve high forecasting accuracy by learning complex temporal patterns; however, their extensive parameters and substantial computational costs make practical deployment difficult in latency-sensitive environments. Therefore, lightweight models based on linear layers have recently been studied for improved efficiency. However, existing linear-based models have difficulty capturing local patterns and fail to reflect sudden volatility or fine-grained local trends, limiting their overall representational capacity. In this paper, SegTSF is proposed, a linear-layer-based lightweight model for multivariate time-series forecasting that improves forecasting performance by enhancing the representational capacity of linear layers while maintaining computational efficiency. First, SegTSF reconstructs the input time series into several subsequences in units of periods and explicitly models intra-period relationships with a linear layer to capture detailed temporal information. Second, it divides each subsequence into segment units and applies individual linear layers to the relationships within and between segments to capture local patterns and global trends. Third, each predicted subsequence is reconstructed into its original dimensions to complete the final forecast. Experimental results on benchmark datasets show that the proposed SegTSF achieves performance improvement compared with existing lightweight models while using fewer parameters in various environments. The findings of this study show that SegTSF achieves a balance between efficiency and forecasting performance through hierarchical segment-wise learning within a lightweight architecture.

---

Keywords

---